Human behavior prediction is a key component to studying the spread of wellness and healthy behavior in a social network. In this paper, we introduce an ontology-based Restricted Boltzmann Machine (ORBM) model for human behavior prediction in health social networks. We first propose a bottom-up algorithm to learn the user representation from ontologies. Then the user representation is used to incorporate self-motivation, social inuences, and environmental events together in a human behavior prediction model, which extends a well-known deep learning method, Restricted Boltzmann Machines (RBMs), so that the interactions among the behavior determinants are naturally simulated through parameters. To our best knowledge, the ORBM model is the first ontology-based deep learning approach in health informatics for human behavior prediction. Experiments conducted on both real and synthetic data from health social networks have shown the tremendous effectiveness of our approach compared with conventional methods.